The open nature of collaborative recommender systems present a security problem. Attackers that cannot be readily distinguished from ordinary users may inject biased profiles, degrading the objectivity and accuracy of the system over time. The standard k-nearest neighbor collaborative filtering algorithm has been shown to be quite vulnerable to such attacks. In this paper, we examine extensions to the standard algorithm that supplement similarity weighting of neighbors with a more generic relevance measure. In particular, we consider two techniques, significance weighting and trust weighting, that attempt to calculate the utility of a neighbor with respect to rating prediction. Similar techniques have been used to improve prediction accuracy in collaborative filtering. We show, however, that significance weighting, in particular, results in improved robustness under profile injection attacks, while at the same time providing better recommendation accuracy than both standard k-nearest neighbor approach as well as the trust-based model.
展开▼
